| Age | Commit message (Collapse) | Author |
|
/var/log/guix-containers was a #define, and the daemon read no environment
variables at all, so a host whose /var/log is read-only -- or anyone running
the daemon outside a system-wide install -- had no way to keep container boot
output except to edit and rebuild the C. It is now a runtime value: the flag
wins, then $GUIX_CONTAINER_LOG_DIR, then the old default, and all three uses
(the mkdir, the .writable probe, the per-container <name>.log path) go through
it.
The chosen directory is named at startup ("container logs: DIR") before the
privilege check, for two reasons: it is the first thing an operator needs when
a container does not come up, and it makes the switch observable without root,
which is how test/run.sh now checks it -- --log-dir and the env var each have
to appear in that line, and an unknown option still has to print the usage.
README: documents the flag and the variable, says which wins, and points
ProtectSystem=strict users at --log-dir as the alternative to LogsDirectory=.
Also generalises one line that said "uid=1000(wrk)" -- true of the author's
box, misleading on anyone else's -- to "your own uid, e.g. uid=1000".
|
|
The README explained the protocol, the fingerprint, the privileges and the
privilege boundary well; what it did not have was the sequence for the case
people actually hit -- a container whose services the host then talks to.
New section "Worked example: a service container other programs talk to":
--network shares the host's network namespace so container services are at
127.0.0.1:PORT on the host (and why that is not the daemon's --port); build
the launcher with `guix system container -d` + `guix build` before letting the
daemon start it, because `guix system container` builds *and* runs in the
foreground; and the trap that bit during a real test -- a `--share` whose host
path does not exist is dropped silently, so create it first. Plus the two
timings worth having: a container is at its services ~2 s after START when the
store paths are built, ~45 s on first boot while activation generates keys and
lays out spool directories, so a client that gives up after a few seconds is
measuring the wrong case. Boot output is where to look: LOG/--log, or
/var/log/guix-containers/<pid>.log subject to the unit's LogsDirectory.
LICENSE is now BSD 3-Clause (Copyright (c) 2026 Pjotr Prins), the same text as
ai-memory's, and the README ends with a pointer to it.
|
|
Started as a normal user the daemon happily answered "OK <pid>" to
every START and then forked a launcher that died within milliseconds
(EPERM writing /proc/PID/uid_map), leaving the client with an empty
LIST and, because /var/log was not writable either, no log to explain
it.
Check privileges at startup instead of warning: euid must be 0 and
CapEff must carry CAP_SETUID (the 65536-entry uid map of the
container's user namespace) and CAP_SYS_ADMIN (mounting its file
systems, /proc and /dev); exit 1 with an error naming the missing
piece -- this also catches root under a narrowed
CapabilityBoundingSet=.
Container boot output goes to /var/log/guix-containers: warn when that
directory cannot be created or written (read-only /var, hardened unit)
so the missing log is noticed at start-up rather than while debugging a
failed boot. The hardened systemd example in the README was the cause
of exactly that -- ProtectSystem=strict makes /var read-only -- so it
now sets LogsDirectory=guix-containers, and a new "Privileges required"
section documents the requirement.
test/test.c gains test_privileges() (CapEff parse, root/non-root
verdict); ./test/run.sh still passes.
|
|
Document the privilege boundary that motivates the daemon (guix system
container needs real root: namespaces, CAP_NET_ADMIN, root shepherd;
AppArmor fixes do not change that) and the verified rootless fallback:
guix shell -C works from an unprivileged account when user namespaces
are allowed and an explicit writable profile path is used (-p); no
shepherd/sshd/guix-system services inside, but foreground daemons
(nginx/php-fpm/octave) run fine for smoke tests. Also note the
unprivileged guix-daemon (builds only), rootless Podman and PRoot
alternatives.
|
|
|
|
VM definitions are not services; give them their own directory.
The ../etc key reference still resolves from vms/.
|
|
Replace (cons hello %base-packages) with an explicit minimal list:
bash, coreutils, findutils, grep, sed, gawk, tar, gzip, shadow,
procps, guile and hello. %base-packages would add ~400 binaries to
/run/current-system/profile/bin (nano, man-db, sudo, the inetutils
ftp/telnet, wget, pciutils, ...). The ssh tools still come in via
the openssh service and 'herd' via the essential shepherd service;
the (useless) guix service is deleted from %base-services like in
the containers.
Also fix a list-building bug: (cons* ... hello) makes hello the
TAIL of the list, so use 'list'.
Result: profile/bin has 185 binaries (was 394) and the image store
holds 263 items (was 362). Verified: ssh login, hello, and
'herd status ssh-daemon' all work.
|
|
Unlike 'guix system container', which bind-mounts the WHOLE host
/gnu/store read-only into the guest (hardcoded %store-mapping in
gnu/system/linux-container.scm -- there is no supported way to mount
a subset), a qcow2 disk image built with 'guix system image -t qcow2'
contains only the closure of the system: the guest sees ~360 store
items instead of the host's 51815 and no host file systems at all.
The definition is the minimal hello+ssh OS with %qemu-static-networking
(QEMU user-mode network stack) and a serial console on ttyS0. Host
SSH keys are generated at first boot and persist in the writable
image, so they need no pinning.
Verified: booted under qemu-system-x86_64 (TCG), ssh on forwarded
port 2223 works with etc/hello-container-key, 'hello' runs, and the
guest /gnu/store lists 362 items.
|
|
|
|
Containers that hang at boot (e.g. blocked on entropy) ignore
SIGTERM; kill -9 them instead.
|
|
* START is refused while a previously started container is still
alive (ERR already running container <pid>)
* new LIST command reports running containers, including those
started by previous daemon instances (via a /proc scan), so they
can be found and KILLed after a daemon restart
* client: new --list option; README updated
|
|
A C daemon (run as root) that starts Guix system containers on
request over a line-based network protocol, plus an unprivileged
Ruby client.
* START/KILL/LOG/FINGERPRINT protocol; one request per connection
* strict validation of store items before exec (HASH-run-container
name, guile shebang, container markers, immutable store file)
* --fingerprint WORD identifies the daemon instance (random hash
when unset); verified by clients and passed to containers in
GUIX_CONTAINER_FINGERPRINT
* on SIGINT/SIGTERM: SIGTERM all started containers, 10s timeout,
then SIGKILL
* container boot output appended to /var/log/guix-containers/
* unit tests for launcher and fingerprint logic
|